On 16 March 2016 at 16:23, Tomeu Vizoso wrote: > On 15 March 2016 at 02:30, Mark yao wrote: >> On 2016年03月14日 21:35, Tomeu Vizoso wrote: >>> >>> On 2 December 2014 at 10:15, Mark Yao wrote: >>>> >>>> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c >>>> b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c >>>> new file mode 100644 >>>> index 0000000..e7ca25b >>>> --- /dev/null >>>> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c >>>> @@ -0,0 +1,1455 @@ >>> >>> ... >>>> >>>> +static bool vop_crtc_mode_fixup(struct drm_crtc *crtc, >>>> + const struct drm_display_mode *mode, >>>> + struct drm_display_mode *adjusted_mode) >>>> +{ >>>> + if (adjusted_mode->htotal == 0 || adjusted_mode->vtotal == 0) >>>> + return false; >>> >>> Hi Mark, >>> >>> what's the rationale for this? >>> >>> Disabling a CRTC as in [0] will cause mode_fixup() to be called with >>> an empty mode, failing that test. >>> >>> Removing the check seems to get things working fine for a short while, >>> but a later modeset invariably gets the VOP to hang (as reported by >>> [1]). >>> >>> Do you know why that check was put in place and what exactly could be >>> causing the hw to hang? >>> >>> [0] >>> https://cgit.freedesktop.org/xorg/app/intel-gpu-tools/tree/lib/igt_kms.c#n1616 >>> [1] >>> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/tree/drivers/gpu/drm/rockchip/rockchip_drm_vop.c#n873 >>> >>> Thanks, >>> >>> Tomeu >>> >> Hi Tomeu >> >> Just thinking that "adjusted_mode->htotal == 0 || adjusted_mode->vtotal == >> 0" is not a good mode for vop. > > Ah, ok. Guess it should be removed then so we don't break userspace? > >> And you said VOP hang, only WARN_ON error message? or system hang, die? > > Sorry, the symptom was only the warning, I just went a bit too far and > assumed the VOP had stopped working at all. > >> I think maybe crtc disable too fast, vblank is off, then no one can feed the >> wait_update_complete. >> Can you test it again with following patch? > > Actually, in today's testing I don't see that happening any more, > sorry about that :/ > > What I have been looking at today is a related issue when running the > kms_flip_event_leak test from intel-gpu-tools. If I remove the check > mentioned above so CRTCs can be disabled with the SetCRTC IOCTL, I see > this page fault the second and subsequent times I run the test. > > [ 75.809031] rk_iommu ff930300.iommu: Page fault at 0x01000000 of type read > [ 75.809035] rk_iommu ff930300.iommu: iova = 0x01000000: dte_index: > 0x4 pte_index: 0x0 page_offset: 0x0 > [ 75.809040] rk_iommu ff930300.iommu: mmu_dte_addr: 0x2c258000 > dte@0x2c258010: 0x2c561001 valid: 1 pte@0x2c561000: 0x2a000006 valid: > 0 page@0x00000000 flags: 0x0 > [ 76.951288] rk_iommu ff930300.iommu: Enable stall request timed > out, status: 0x00004b > > I have written a smaller standalone test that is attached in case you > want to check it out, but I haven't been able to find out why it only > happens when the test is rerun. > > Apparently the VOP is still trying to read a BO (0x01000000) right > when the kernel frees it, but from what I can see, it should be > scanning another BO at that point. > > Do you have any ideas on what could be happening? Apparently, when the VOP is re-enabled, it will start scanning from the framebuffer address that had been set last. Because DMA addresses are recycled and there's going to be a low number of framebuffers, this isn't going to be obvious unless we make sure that there isn't a FB allocated at that DMA address any more. The attached test case does just that. Regards, Tomeu > Thanks, > > Tomeu > >> drivers/gpu/drm/rockchip/rockchip_drm_vop.c >> @@ -503,6 +503,8 @@ static void vop_crtc_disable(struct drm_crtc *crtc) >> if (!vop->is_enabled) >> return; >> + vop_crtc_wait_for_update(crtc); >> + >> drm_crtc_vblank_off(crtc); >> >> Thanks. >> >> -- >> Ｍark Yao >> >>